[saketh]

For a case of 1 million cells in fluent, RAM consumption is less than 150 mb where as in our solver it is 12 gb. what is the key thing that we are missing in our data structure?

[selig5576]

Hi, what is probably missing is the fact that you are not doing memory allocation. If you are doing Fortran or C++ use dynamic arrays (or better yet if you are doing C++ use vectors.)

[sbaffini]

That seems extremely low even for a 2D computation. Are you sure it is not 1.5 Gb? Because 150 Mb is just around 18 double precision floating point values per cell. Take at least 8 of them (4 variables in 3D + cell volume + 3 for cell centroid). Then consider that faces are typically 3 times the number of cells, and you need at least the face area vector (3 doubles), the face centroid (3 doubles) and the mass flux (1 double), totalling 7 x 3 = 21 doubles per cell. We already are at 29 doubles and I'm not even considering the grid nodes (3 more doubles per cell approx.), one double to store at least the current residual, one to store the current solution increment. Then you have all sort of indexing for the internal machinery. The matrix coefficients (say, approx 7 doubles per cell), etc. etc.

In contrast, your 12 Gb quote is quite high and might be due to lack of deallocation. Or, maybe, you are using more variables than needed (e.g., for a segregated solver, you only need one matrix and one solution vector).

This might be ok early in the development phase, but then you need to figure out more clever ways to retrieve some information without storing them directly.

If you give more details on your code we might help you more.

[nkalkote]

............